xref: /illumos-gate/usr/src/cmd/stat/iostat/iostat.c (revision 48bbca816818409505a6e214d0911fda44e622e3)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  *
26  * rewritten from UCB 4.13 83/09/25
27  * rewritten from SunOS 4.1 SID 1.18 89/10/06
28  */
29 /*
30  * Copyright (c) 2012 by Delphix. All rights reserved.
31  * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
32  * Copyright 2016 James S. Blachly, MD. All rights reserved.
33  */
34 
35 #include <stdio.h>
36 #include <stdlib.h>
37 #include <stdarg.h>
38 #include <ctype.h>
39 #include <unistd.h>
40 #include <memory.h>
41 #include <errno.h>
42 #include <string.h>
43 #include <signal.h>
44 #include <sys/types.h>
45 #include <time.h>
46 #include <sys/time.h>
47 #include <sys/sysinfo.h>
48 #include <inttypes.h>
49 #include <strings.h>
50 #include <sys/systeminfo.h>
51 #include <kstat.h>
52 #include <locale.h>
53 
54 #include "dsr.h"
55 #include "statcommon.h"
56 
57 #define	DISK_OLD		0x0001
58 #define	DISK_NEW		0x0002
59 #define	DISK_EXTENDED		0x0004
60 #define	DISK_ERRORS		0x0008
61 #define	DISK_EXTENDED_ERRORS	0x0010
62 #define	DISK_IOPATH_LI		0x0020	/* LunInitiator */
63 #define	DISK_IOPATH_LTI		0x0040	/* LunTargetInitiator */
64 
65 #define	DISK_NORMAL		(DISK_OLD | DISK_NEW)
66 #define	DISK_IO_MASK		(DISK_OLD | DISK_NEW | DISK_EXTENDED)
67 #define	DISK_ERROR_MASK		(DISK_ERRORS | DISK_EXTENDED_ERRORS)
68 #define	PRINT_VERTICAL		(DISK_ERROR_MASK | DISK_EXTENDED)
69 
70 #define	REPRINT 19
71 
72 #define	NUMBER_OF_ERR_COUNTERS	3
73 
74 /*
75  * It's really a pseudo-gigabyte. We use 1000000000 bytes so that the disk
76  * labels don't look bad. 1GB is really 1073741824 bytes.
77  */
78 #define	DISK_GIGABYTE   1000000000.0
79 
80 /*
81  * Function desciptor to be called when extended
82  * headers are used.
83  */
84 typedef struct formatter {
85 	void (*nfunc)(void);
86 	struct formatter *next;
87 } format_t;
88 
89 /*
90  * Used to get formatting right when printing tty/cpu
91  * data to the right of disk data
92  */
93 enum show_disk_mode {
94 	SHOW_FIRST_ONLY,
95 	SHOW_SECOND_ONWARDS,
96 	SHOW_ALL
97 };
98 
99 enum show_disk_mode show_disk_mode = SHOW_ALL;
100 
101 char *cmdname = "iostat";
102 int caught_cont = 0;
103 
104 static char one_blank[] = " ";
105 static char two_blanks[] = "  ";
106 
107 /*
108  * count for number of lines to be emitted before a header is
109  * shown again. Only used for the basic format.
110  */
111 static	uint_t	tohdr = 1;
112 
113 /*
114  * If we're in raw format, have we printed a header? We only do it
115  * once for raw but we emit it every REPRINT lines in non-raw format.
116  * This applies only for the basic header. The extended header is
117  * done only once in both formats.
118  */
119 static	uint_t	hdr_out;
120 
121 /*
122  * Flags representing arguments from command line
123  */
124 static	uint_t	do_tty;			/* show tty info (-t) */
125 static	uint_t	do_disk;		/* show disk info per selected */
126 					/* format (-d, -D, -e, -E, -x -X -Y) */
127 static	uint_t	do_cpu;			/* show cpu info (-c) */
128 static	uint_t	do_interval;		/* do intervals (-I) */
129 static	int	do_partitions;		/* per-partition stats (-p) */
130 static	int	do_partitions_only;	/* per-partition stats only (-P) */
131 					/* no per-device stats for disks */
132 static	uint_t	do_conversions;		/* display disks as cXtYdZ (-n) */
133 static	uint_t	do_megabytes;		/* display data in MB/sec (-M) */
134 static  uint_t	do_controller;		/* display controller info (-C) */
135 static  uint_t	do_raw;			/* emit raw format (-r) */
136 static	uint_t	timestamp_fmt = NODATE;	/* timestamp  each display (-T) */
137 static	uint_t	do_devid;		/* -E should show devid */
138 
139 /*
140  * Default number of disk drives to be displayed in basic format
141  */
142 #define	DEFAULT_LIMIT	4
143 
144 struct iodev_filter df;
145 
146 static  uint_t	suppress_state;		/* skip state change messages */
147 static	uint_t	suppress_zero;		/* skip zero valued lines */
148 static  uint_t	show_mountpts;		/* show mount points */
149 static	int 	interval;		/* interval (seconds) to output */
150 static	int 	iter;			/* iterations from command line */
151 
152 #define	SMALL_SCRATCH_BUFLEN	MAXNAMELEN
153 
154 static int	iodevs_nl;		/* name field width */
155 #define	IODEVS_NL_MIN		6	/* not too thin for "device" */
156 #define	IODEVS_NL_MAX		24	/* but keep full width under 80 */
157 
158 static	char	disk_header[132];
159 static	uint_t 	dh_len;			/* disk header length for centering */
160 static  int 	lineout;		/* data waiting to be printed? */
161 
162 static struct snapshot *newss;
163 static struct snapshot *oldss;
164 static	double	getime;			/* elapsed time */
165 static	double	percent;		/* 100 / etime */
166 
167 /*
168  * List of functions to be called which will construct the desired output
169  */
170 static format_t	*formatter_list;
171 static format_t *formatter_end;
172 
173 static u_longlong_t	ull_delta(u_longlong_t, u_longlong_t);
174 static uint_t 	u32_delta(uint_t, uint_t);
175 static void setup(void (*nfunc)(void));
176 static void print_tty_hdr1(void);
177 static void print_tty_hdr2(void);
178 static void print_cpu_hdr1(void);
179 static void print_cpu_hdr2(void);
180 static void print_tty_data(void);
181 static void print_cpu_data(void);
182 static void print_err_hdr(void);
183 static void print_disk_header(void);
184 static void hdrout(void);
185 static void disk_errors(void);
186 static void do_newline(void);
187 static void push_out(const char *, ...);
188 static void printhdr(int);
189 static void printxhdr(void);
190 static void usage(void);
191 static void do_args(int, char **);
192 static void do_format(void);
193 static void show_all_disks(void);
194 static void show_first_disk(void);
195 static void show_other_disks(void);
196 static void show_disk_errors(void *, void *, void *);
197 static void write_core_header(void);
198 static int  fzero(double value);
199 static int  safe_strtoi(char const *val, char *errmsg);
200 
201 int
202 main(int argc, char **argv)
203 {
204 	enum snapshot_types types = SNAP_SYSTEM;
205 	kstat_ctl_t *kc;
206 	long hz;
207 	int forever;
208 	hrtime_t start_n;
209 	hrtime_t period_n;
210 
211 	(void) setlocale(LC_ALL, "");
212 #if !defined(TEXT_DOMAIN)		/* Should be defined by cc -D */
213 #define	TEXT_DOMAIN "SYS_TEST"		/* Use this only if it weren't */
214 #endif
215 	(void) textdomain(TEXT_DOMAIN);
216 
217 	do_args(argc, argv);
218 
219 	/*
220 	 * iostat historically showed CPU changes, even though
221 	 * it doesn't provide much useful information
222 	 */
223 	types |= SNAP_CPUS;
224 
225 	if (do_disk)
226 		types |= SNAP_IODEVS;
227 
228 	if (do_disk && !do_partitions_only)
229 		df.if_allowed_types |= IODEV_DISK;
230 	if (do_disk & DISK_IOPATH_LI) {
231 		df.if_allowed_types |= IODEV_IOPATH_LTI;
232 		types |= SNAP_IOPATHS_LI;
233 	}
234 	if (do_disk & DISK_IOPATH_LTI) {
235 		df.if_allowed_types |= IODEV_IOPATH_LTI;
236 		types |= SNAP_IOPATHS_LTI;
237 	}
238 	if (do_disk & DISK_ERROR_MASK)
239 		types |= SNAP_IODEV_ERRORS;
240 	if (do_partitions || do_partitions_only)
241 		df.if_allowed_types |= IODEV_PARTITION;
242 	if (do_conversions)
243 		types |= SNAP_IODEV_PRETTY;
244 	if (do_devid)
245 		types |= SNAP_IODEV_DEVID;
246 	if (do_controller) {
247 		if (!(do_disk & PRINT_VERTICAL) ||
248 		    (do_disk & DISK_EXTENDED_ERRORS))
249 			fail(0, "-C can only be used with -e or -x.");
250 		types |= SNAP_CONTROLLERS;
251 		df.if_allowed_types |= IODEV_CONTROLLER;
252 	}
253 
254 	hz = sysconf(_SC_CLK_TCK);
255 
256 	/*
257 	 * Undocumented behavior - sending a SIGCONT will result
258 	 * in a new header being emitted. Used only if we're not
259 	 * doing extended headers. This is a historical
260 	 * artifact.
261 	 */
262 	if (!(do_disk & PRINT_VERTICAL))
263 		(void) signal(SIGCONT, printhdr);
264 
265 	if (interval)
266 		period_n = (hrtime_t)interval * NANOSEC;
267 
268 	kc = open_kstat();
269 	if (interval)
270 		start_n = gethrtime();
271 	newss = acquire_snapshot(kc, types, &df);
272 
273 	/* compute width of "device" field */
274 	iodevs_nl = newss->s_iodevs_is_name_maxlen;
275 	iodevs_nl = (iodevs_nl < IODEVS_NL_MIN) ?
276 	    IODEVS_NL_MIN : iodevs_nl;
277 	iodevs_nl = (iodevs_nl > IODEVS_NL_MAX) ?
278 	    IODEVS_NL_MAX : iodevs_nl;
279 
280 	do_format();
281 
282 	forever = (iter == 0);
283 	do {
284 		if (do_conversions && show_mountpts)
285 			do_mnttab();
286 
287 		if (do_tty || do_cpu) {
288 			kstat_t *oldks;
289 			oldks = oldss ? &oldss->s_sys.ss_agg_sys : NULL;
290 			getime = cpu_ticks_delta(oldks,
291 			    &newss->s_sys.ss_agg_sys);
292 			percent = (getime > 0.0) ? 100.0 / getime : 0.0;
293 			getime = (getime / nr_active_cpus(newss)) / hz;
294 			if (getime == 0.0)
295 				getime = (double)interval;
296 			if (getime == 0.0 || do_interval)
297 				getime = 1.0;
298 		}
299 
300 		if (formatter_list) {
301 			format_t *tmp;
302 			tmp = formatter_list;
303 
304 			if (timestamp_fmt != NODATE)
305 				print_timestamp(timestamp_fmt);
306 
307 			while (tmp) {
308 				(tmp->nfunc)();
309 				tmp = tmp->next;
310 			}
311 			(void) fflush(stdout);
312 		}
313 
314 		/* only remaining/doing a single iteration, we are done */
315 		if (iter == 1)
316 			continue;
317 
318 		if (interval > 0)
319 			/* Have a kip */
320 			sleep_until(&start_n, period_n, forever, &caught_cont);
321 
322 		free_snapshot(oldss);
323 		oldss = newss;
324 		newss = acquire_snapshot(kc, types, &df);
325 		iodevs_nl = (newss->s_iodevs_is_name_maxlen > iodevs_nl) ?
326 		    newss->s_iodevs_is_name_maxlen : iodevs_nl;
327 		iodevs_nl = (iodevs_nl < IODEVS_NL_MIN) ?
328 		    IODEVS_NL_MIN : iodevs_nl;
329 		iodevs_nl = (iodevs_nl > IODEVS_NL_MAX) ?
330 		    IODEVS_NL_MAX : iodevs_nl;
331 
332 		if (!suppress_state)
333 			snapshot_report_changes(oldss, newss);
334 
335 		/* if config changed, show stats from boot */
336 		if (snapshot_has_changed(oldss, newss)) {
337 			free_snapshot(oldss);
338 			oldss = NULL;
339 		}
340 
341 	} while (--iter);
342 
343 	free_snapshot(oldss);
344 	free_snapshot(newss);
345 	(void) kstat_close(kc);
346 	free(df.if_names);
347 	return (0);
348 }
349 
350 /*
351  * Some magic numbers used in header formatting.
352  *
353  * DISK_LEN = length of either "kps tps serv" or "wps rps util"
354  *	      using 0 as the first position
355  *
356  * DISK_ERROR_LEN = length of "s/w h/w trn tot" with one space on
357  *		either side. Does not use zero as first pos.
358  *
359  * DEVICE_LEN = length of "device" + 1 character.
360  */
361 
362 #define	DISK_LEN	11
363 #define	DISK_ERROR_LEN	16
364 #define	DEVICE_LEN	7
365 
366 /*ARGSUSED*/
367 static void
368 show_disk_name(void *v1, void *v2, void *data)
369 {
370 	struct iodev_snapshot *dev = (struct iodev_snapshot *)v2;
371 	size_t slen;
372 	char *name;
373 	char fbuf[SMALL_SCRATCH_BUFLEN];
374 
375 	if (dev == NULL)
376 		return;
377 
378 	name = do_conversions ? dev->is_pretty : dev->is_name;
379 	name = name ? name : dev->is_name;
380 
381 	if (!do_raw) {
382 		uint_t width;
383 
384 		slen = strlen(name);
385 		/*
386 		 * The length is less
387 		 * than the section
388 		 * which will be displayed
389 		 * on the next line.
390 		 * Center the entry.
391 		 */
392 
393 		width = (DISK_LEN + 1)/2 + (slen / 2);
394 		(void) snprintf(fbuf, sizeof (fbuf),
395 		    "%*s", width, name);
396 		name = fbuf;
397 		push_out("%-13.13s ", name);
398 	} else {
399 		push_out(name);
400 	}
401 }
402 
403 /*ARGSUSED*/
404 static void
405 show_disk_header(void *v1, void *v2, void *data)
406 {
407 	push_out(disk_header);
408 }
409 
410 /*
411  * Write out a two line header. What is written out depends on the flags
412  * selected but in the worst case consists of a tty header, a disk header
413  * providing information for 4 disks and a cpu header.
414  *
415  * The tty header consists of the word "tty" on the first line above the
416  * words "tin tout" on the next line. If present the tty portion consumes
417  * the first 10 characters of each line since "tin tout" is surrounded
418  * by single spaces.
419  *
420  * Each of the disk sections is a 14 character "block" in which the name of
421  * the disk is centered in the first 12 characters of the first line.
422  *
423  * The cpu section is an 11 character block with "cpu" centered over the
424  * section.
425  *
426  * The worst case should look as follows:
427  *
428  * 0---------1--------2---------3---------4---------5---------6---------7-------
429  *    tty        sd0           sd1           sd2           sd3           cpu
430  *  tin tout kps tps serv  kps tps serv  kps tps serv  kps tps serv  us sy dt id
431  *  NNN NNNN NNN NNN NNNN  NNN NNN NNNN  NNN NNN NNNN  NNN NNN NNNN  NN NN NN NN
432  *
433  * When -D is specified, the disk header looks as follows (worst case):
434  *
435  * 0---------1--------2---------3---------4---------5---------6---------7-------
436  *     tty        sd0           sd1             sd2          sd3          cpu
437  *   tin tout rps wps util  rps wps util  rps wps util  rps wps util us sy dt id
438  *   NNN NNNN NNN NNN NNNN  NNN NNN NNNN  NNN NNN NNNN  NNN NNN NNNN NN NN NN NN
439  */
440 static void
441 printhdr(int sig)
442 {
443 	/*
444 	 * If we're here because a signal fired, reenable the
445 	 * signal.
446 	 */
447 	if (sig)
448 		(void) signal(SIGCONT, printhdr);
449 	if (sig == SIGCONT)
450 		caught_cont = 1;
451 	/*
452 	 * Horizontal mode headers
453 	 *
454 	 * First line
455 	 */
456 	if (do_tty)
457 		print_tty_hdr1();
458 
459 	if (do_disk & DISK_NORMAL) {
460 		(void) snapshot_walk(SNAP_IODEVS, NULL, newss,
461 		    show_disk_name, NULL);
462 	}
463 
464 	if (do_cpu)
465 		print_cpu_hdr1();
466 	do_newline();
467 
468 	/*
469 	 * Second line
470 	 */
471 	if (do_tty)
472 		print_tty_hdr2();
473 
474 	if (do_disk & DISK_NORMAL) {
475 		(void) snapshot_walk(SNAP_IODEVS, NULL, newss,
476 		    show_disk_header, NULL);
477 	}
478 
479 	if (do_cpu)
480 		print_cpu_hdr2();
481 	do_newline();
482 
483 	tohdr = REPRINT;
484 }
485 
486 /*
487  * Write out the extended header centered over the core information.
488  */
489 static void
490 write_core_header(void)
491 {
492 	char *edev = "extended device statistics";
493 	uint_t lead_space_ct;
494 	uint_t follow_space_ct;
495 	size_t edevlen;
496 
497 	if (do_raw == 0) {
498 		/*
499 		 * The things we do to look nice...
500 		 *
501 		 * Center the core output header. Make sure we have the
502 		 * right number of trailing spaces for follow-on headers
503 		 * (i.e., cpu and/or tty and/or errors).
504 		 */
505 		edevlen = strlen(edev);
506 		lead_space_ct = dh_len - edevlen;
507 		lead_space_ct /= 2;
508 		if (lead_space_ct > 0) {
509 			follow_space_ct = dh_len - (lead_space_ct + edevlen);
510 			if (do_disk & DISK_ERRORS)
511 				follow_space_ct -= DISK_ERROR_LEN;
512 			if ((do_disk & DISK_EXTENDED) && do_conversions)
513 				follow_space_ct -= DEVICE_LEN;
514 
515 			push_out("%1$*2$.*2$s%3$s%4$*5$.*5$s", one_blank,
516 			    lead_space_ct, edev, one_blank, follow_space_ct);
517 		} else
518 			push_out("%56s", edev);
519 	} else
520 		push_out(edev);
521 }
522 
523 /*
524  * In extended mode headers, we don't want to reprint the header on
525  * signals as they are printed every time anyways.
526  */
527 static void
528 printxhdr(void)
529 {
530 
531 	/*
532 	 * Vertical mode headers
533 	 */
534 	if (do_disk & DISK_EXTENDED)
535 		setup(write_core_header);
536 	if (do_disk & DISK_ERRORS)
537 		setup(print_err_hdr);
538 
539 	if (do_conversions) {
540 		setup(do_newline);
541 		if (do_disk & (DISK_EXTENDED | DISK_ERRORS))
542 			setup(print_disk_header);
543 		setup(do_newline);
544 	} else {
545 		if (do_tty)
546 			setup(print_tty_hdr1);
547 		if (do_cpu)
548 			setup(print_cpu_hdr1);
549 		setup(do_newline);
550 
551 		if (do_disk & (DISK_EXTENDED | DISK_ERRORS))
552 			setup(print_disk_header);
553 		if (do_tty)
554 			setup(print_tty_hdr2);
555 		if (do_cpu)
556 			setup(print_cpu_hdr2);
557 		setup(do_newline);
558 	}
559 }
560 
561 /*
562  * Write out a line for this disk - note that show_disk writes out
563  * full lines or blocks for each selected disk.
564  */
565 static void
566 show_disk(void *v1, void *v2, void *data)
567 {
568 	uint32_t err_counters[NUMBER_OF_ERR_COUNTERS];
569 	boolean_t display_err_counters = do_disk & DISK_ERRORS;
570 	struct iodev_snapshot *old = (struct iodev_snapshot *)v1;
571 	struct iodev_snapshot *new = (struct iodev_snapshot *)v2;
572 	int *count = (int *)data;
573 	double rps, wps, tps, mtps, krps, kwps, kps, avw, avr, w_pct, r_pct;
574 	double wserv, rserv, serv;
575 	double iosize;	/* kb/sec or MB/sec */
576 	double etime, hr_etime;
577 	char *disk_name;
578 	u_longlong_t ldeltas;
579 	uint_t udeltas;
580 	uint64_t t_delta;
581 	uint64_t w_delta;
582 	uint64_t r_delta;
583 	int doit = 1;
584 	uint_t toterrs;
585 	char *fstr;
586 
587 	if (new == NULL)
588 		return;
589 
590 	switch (show_disk_mode) {
591 	case SHOW_FIRST_ONLY:
592 		if (count != NULL && *count)
593 			return;
594 		break;
595 
596 	case SHOW_SECOND_ONWARDS:
597 		if (count != NULL && !*count) {
598 			(*count)++;
599 			return;
600 		}
601 		break;
602 
603 	default:
604 		break;
605 	}
606 
607 	disk_name = do_conversions ? new->is_pretty : new->is_name;
608 	disk_name = disk_name ? disk_name : new->is_name;
609 
610 	/*
611 	 * Only do if we want IO stats - Avoids errors traveling this
612 	 * section if that's all we want to see.
613 	 */
614 	if (do_disk & DISK_IO_MASK) {
615 		if (old) {
616 			t_delta = hrtime_delta(old->is_snaptime,
617 			    new->is_snaptime);
618 		} else {
619 			t_delta = hrtime_delta(new->is_crtime,
620 			    new->is_snaptime);
621 		}
622 
623 		if (new->is_nr_children) {
624 			if (new->is_type == IODEV_CONTROLLER) {
625 				t_delta /= new->is_nr_children;
626 			} else if ((new->is_type == IODEV_IOPATH_LT) ||
627 			    (new->is_type == IODEV_IOPATH_LI)) {
628 				/* synthetic path */
629 				if (!old) {
630 					t_delta = new->is_crtime;
631 				}
632 				t_delta /= new->is_nr_children;
633 			}
634 		}
635 
636 		hr_etime = (double)t_delta;
637 		if (hr_etime == 0.0)
638 			hr_etime = (double)NANOSEC;
639 		etime = hr_etime / (double)NANOSEC;
640 
641 		/* reads per second */
642 		udeltas = u32_delta(old ? old->is_stats.reads : 0,
643 		    new->is_stats.reads);
644 		rps = (double)udeltas;
645 		rps /= etime;
646 
647 		/* writes per second */
648 		udeltas = u32_delta(old ? old->is_stats.writes : 0,
649 		    new->is_stats.writes);
650 		wps = (double)udeltas;
651 		wps /= etime;
652 
653 		tps = rps + wps;
654 			/* transactions per second */
655 
656 		/*
657 		 * report throughput as either kb/sec or MB/sec
658 		 */
659 
660 		if (!do_megabytes)
661 			iosize = 1024.0;
662 		else
663 			iosize = 1048576.0;
664 
665 		ldeltas = ull_delta(old ? old->is_stats.nread : 0,
666 		    new->is_stats.nread);
667 		if (ldeltas) {
668 			krps = (double)ldeltas;
669 			krps /= etime;
670 			krps /= iosize;
671 		} else
672 			krps = 0.0;
673 
674 		ldeltas = ull_delta(old ? old->is_stats.nwritten : 0,
675 		    new->is_stats.nwritten);
676 		if (ldeltas) {
677 			kwps = (double)ldeltas;
678 			kwps /= etime;
679 			kwps /= iosize;
680 		} else
681 			kwps = 0.0;
682 
683 		/*
684 		 * Blocks transferred per second
685 		 */
686 		kps = krps + kwps;
687 
688 		/*
689 		 * Average number of wait transactions waiting
690 		 */
691 		w_delta = hrtime_delta((u_longlong_t)
692 		    (old ? old->is_stats.wlentime : 0),
693 		    new->is_stats.wlentime);
694 		if (w_delta) {
695 			avw = (double)w_delta;
696 			avw /= hr_etime;
697 		} else
698 			avw = 0.0;
699 
700 		/*
701 		 * Average number of run transactions waiting
702 		 */
703 		r_delta = hrtime_delta(old ? old->is_stats.rlentime : 0,
704 		    new->is_stats.rlentime);
705 		if (r_delta) {
706 			avr = (double)r_delta;
707 			avr /= hr_etime;
708 		} else
709 			avr = 0.0;
710 
711 		/*
712 		 * Average wait service time in milliseconds
713 		 */
714 		if (tps > 0.0 && (avw != 0.0 || avr != 0.0)) {
715 			mtps = 1000.0 / tps;
716 			if (avw != 0.0)
717 				wserv = avw * mtps;
718 			else
719 				wserv = 0.0;
720 
721 			if (avr != 0.0)
722 				rserv = avr * mtps;
723 			else
724 				rserv = 0.0;
725 			serv = rserv + wserv;
726 		} else {
727 			rserv = 0.0;
728 			wserv = 0.0;
729 			serv = 0.0;
730 		}
731 
732 		/* % of time there is a transaction waiting for service */
733 		t_delta = hrtime_delta(old ? old->is_stats.wtime : 0,
734 		    new->is_stats.wtime);
735 		if (t_delta) {
736 			w_pct = (double)t_delta;
737 			w_pct /= hr_etime;
738 			w_pct *= 100.0;
739 
740 			/*
741 			 * Average the wait queue utilization over the
742 			 * the controller's devices, if this is a controller.
743 			 */
744 			if (new->is_type == IODEV_CONTROLLER)
745 				w_pct /= new->is_nr_children;
746 		} else
747 			w_pct = 0.0;
748 
749 		/* % of time there is a transaction running */
750 		t_delta = hrtime_delta(old ? old->is_stats.rtime : 0,
751 		    new->is_stats.rtime);
752 		if (t_delta) {
753 			r_pct = (double)t_delta;
754 			r_pct /= hr_etime;
755 			r_pct *= 100.0;
756 
757 			/*
758 			 * Average the percent busy over the controller's
759 			 * devices, if this is a controller.
760 			 */
761 			if (new->is_type == IODEV_CONTROLLER)
762 				w_pct /= new->is_nr_children;
763 		} else {
764 			r_pct = 0.0;
765 		}
766 
767 		/* % of time there is a transaction running */
768 		if (do_interval) {
769 			rps	*= etime;
770 			wps	*= etime;
771 			tps	*= etime;
772 			krps	*= etime;
773 			kwps	*= etime;
774 			kps	*= etime;
775 		}
776 	}
777 
778 	if (do_disk & (DISK_EXTENDED | DISK_ERRORS)) {
779 		if ((!do_conversions) && ((suppress_zero == 0) ||
780 		    ((do_disk & DISK_EXTENDED) == 0))) {
781 			if (do_raw == 0) {
782 				push_out("%-*.*s",
783 				    iodevs_nl, iodevs_nl, disk_name);
784 			} else {
785 				push_out(disk_name);
786 			}
787 		}
788 	}
789 
790 	/*
791 	 * The error counters are read first (if asked for and if they are
792 	 * available).
793 	 */
794 	bzero(err_counters, sizeof (err_counters));
795 	toterrs = 0;
796 	if (display_err_counters && (new->is_errors.ks_data != NULL)) {
797 		kstat_named_t	*knp;
798 		int		i;
799 
800 		knp = KSTAT_NAMED_PTR(&new->is_errors);
801 		for (i = 0; i < NUMBER_OF_ERR_COUNTERS; i++) {
802 			switch (knp[i].data_type) {
803 				case KSTAT_DATA_ULONG:
804 				case KSTAT_DATA_ULONGLONG:
805 					err_counters[i] = knp[i].value.ui32;
806 					toterrs += knp[i].value.ui32;
807 					break;
808 				default:
809 					break;
810 			}
811 		}
812 	}
813 
814 	switch (do_disk & DISK_IO_MASK) {
815 	case DISK_OLD:
816 		if (do_raw == 0)
817 			fstr = "%3.0f %3.0f %4.0f  ";
818 		else
819 			fstr = "%.0f,%.0f,%.0f";
820 		push_out(fstr, kps, tps, serv);
821 		break;
822 	case DISK_NEW:
823 		if (do_raw == 0)
824 			fstr = "%3.0f %3.0f %4.1f  ";
825 		else
826 			fstr = "%.0f,%.0f,%.1f";
827 		push_out(fstr, rps, wps, r_pct);
828 		break;
829 	case DISK_EXTENDED:
830 		if (suppress_zero) {
831 			if (fzero(rps) && fzero(wps) && fzero(krps) &&
832 			    fzero(kwps) && fzero(avw) && fzero(avr) &&
833 			    fzero(serv) && fzero(w_pct) && fzero(r_pct) &&
834 			    (toterrs == 0)) {
835 				doit = 0;
836 				display_err_counters = B_FALSE;
837 			} else if (do_conversions == 0) {
838 				if (do_raw == 0) {
839 					push_out("%-*.*s",
840 					    iodevs_nl, iodevs_nl, disk_name);
841 				} else {
842 					push_out(disk_name);
843 				}
844 			}
845 		}
846 		if (doit) {
847 			if (!do_conversions) {
848 				if (do_raw == 0) {
849 					fstr = " %6.1f %6.1f %6.1f %6.1f "
850 					    "%4.1f %4.1f %6.1f %3.0f "
851 					    "%3.0f ";
852 				} else {
853 					fstr = "%.1f,%.1f,%.1f,%.1f,%.1f,%.1f,"
854 					    "%.1f,%.0f,%.0f";
855 				}
856 				push_out(fstr, rps, wps, krps, kwps, avw, avr,
857 				    serv, w_pct, r_pct);
858 			} else {
859 				if (do_raw == 0) {
860 					fstr = " %6.1f %6.1f %6.1f %6.1f "
861 					    "%4.1f %4.1f %6.1f %6.1f "
862 					    "%3.0f %3.0f ";
863 				} else {
864 					fstr = "%.1f,%.1f,%.1f,%.1f,%.1f,%.1f,"
865 					    "%.1f,%.1f,%.0f,%.0f";
866 				}
867 				push_out(fstr, rps, wps, krps, kwps, avw, avr,
868 				    wserv, rserv, w_pct, r_pct);
869 			}
870 		}
871 		break;
872 	}
873 
874 	if (display_err_counters) {
875 		char	*efstr;
876 		int	i;
877 
878 		if (do_raw == 0) {
879 			if (do_disk == DISK_ERRORS)
880 				push_out(two_blanks);
881 			efstr = "%3u ";
882 		} else {
883 			efstr = "%u";
884 		}
885 
886 		for (i = 0; i < NUMBER_OF_ERR_COUNTERS; i++)
887 			push_out(efstr, err_counters[i]);
888 
889 		push_out(efstr, toterrs);
890 	}
891 
892 	if (suppress_zero == 0 || doit == 1) {
893 		if ((do_disk & (DISK_EXTENDED | DISK_ERRORS)) &&
894 		    do_conversions) {
895 			push_out("%s", disk_name);
896 			if (show_mountpts && new->is_dname) {
897 				mnt_t *mount_pt;
898 				char *lu;
899 				char *dnlu;
900 				char lub[SMALL_SCRATCH_BUFLEN];
901 
902 				lu = strrchr(new->is_dname, '/');
903 				if (lu) {
904 					/* only the part after a possible '/' */
905 					dnlu = strrchr(disk_name, '/');
906 					if (dnlu != NULL &&
907 					    strcmp(dnlu, lu) == 0)
908 						lu = new->is_dname;
909 					else {
910 						*lu = 0;
911 						(void) strcpy(lub,
912 						    new->is_dname);
913 						*lu = '/';
914 						(void) strcat(lub, "/");
915 						(void) strcat(lub,
916 						    disk_name);
917 						lu = lub;
918 					}
919 				} else
920 					lu = disk_name;
921 				mount_pt = lookup_mntent_byname(lu);
922 				if (mount_pt) {
923 					if (do_raw == 0)
924 						push_out(" (%s)",
925 						    mount_pt->mount_point);
926 					else
927 						push_out("(%s)",
928 						    mount_pt->mount_point);
929 				}
930 			}
931 		}
932 	}
933 
934 	if ((do_disk & PRINT_VERTICAL) && show_disk_mode != SHOW_FIRST_ONLY)
935 		do_newline();
936 
937 	if (count != NULL)
938 		(*count)++;
939 }
940 
941 static void
942 usage(void)
943 {
944 	(void) fprintf(stderr,
945 	    "Usage: iostat [-cCdDeEiImMnpPrstxXYz] "
946 	    " [-l n] [-T d|u] [disk ...] [interval [count]]\n"
947 	    "\t\t-c: 	report percentage of time system has spent\n"
948 	    "\t\t\tin user/system/dtrace/idle mode\n"
949 	    "\t\t-C: 	report disk statistics by controller\n"
950 	    "\t\t-d: 	display disk Kb/sec, transfers/sec, avg. \n"
951 	    "\t\t\tservice time in milliseconds  \n"
952 	    "\t\t-D: 	display disk reads/sec, writes/sec, \n"
953 	    "\t\t\tpercentage disk utilization \n"
954 	    "\t\t-e: 	report device error summary statistics\n"
955 	    "\t\t-E: 	report extended device error statistics\n"
956 	    "\t\t-i:	show device IDs for -E output\n"
957 	    "\t\t-I: 	report the counts in each interval,\n"
958 	    "\t\t\tinstead of rates, where applicable\n"
959 	    "\t\t-l n:	Limit the number of disks to n\n"
960 	    "\t\t-m: 	Display mount points (most useful with -p)\n"
961 	    "\t\t-M: 	Display data throughput in MB/sec "
962 	    "instead of Kb/sec\n"
963 	    "\t\t-n: 	convert device names to cXdYtZ format\n"
964 	    "\t\t-p: 	report per-partition disk statistics\n"
965 	    "\t\t-P: 	report per-partition disk statistics only,\n"
966 	    "\t\t\tno per-device disk statistics\n"
967 	    "\t\t-r: 	Display data in comma separated format\n"
968 	    "\t\t-s: 	Suppress state change messages\n"
969 	    "\t\t-T d|u	Display a timestamp in date (d) or unix "
970 	    "time_t (u)\n"
971 	    "\t\t-t: 	display chars read/written to terminals\n"
972 	    "\t\t-x: 	display extended disk statistics\n"
973 	    "\t\t-X: 	display I/O path statistics\n"
974 	    "\t\t-Y: 	display I/O path (I/T/L) statistics\n"
975 	    "\t\t-z: 	Suppress entries with all zero values\n");
976 	exit(1);
977 }
978 
979 /*ARGSUSED*/
980 static void
981 show_disk_errors(void *v1, void *v2, void *d)
982 {
983 	struct iodev_snapshot *disk = (struct iodev_snapshot *)v2;
984 	kstat_named_t *knp;
985 	size_t  col;
986 	int	i, len;
987 	char	*dev_name;
988 
989 	if (disk->is_errors.ks_ndata == 0)
990 		return;
991 	if (disk->is_type == IODEV_CONTROLLER)
992 		return;
993 
994 	dev_name = do_conversions ? disk->is_pretty : disk->is_name;
995 	dev_name = dev_name ? dev_name : disk->is_name;
996 
997 	len = strlen(dev_name);
998 	if (len > 20)
999 		push_out("%s ", dev_name);
1000 	else if (len > 16)
1001 		push_out("%-20.20s ", dev_name);
1002 	else {
1003 		if (do_conversions)
1004 			push_out("%-16.16s ", dev_name);
1005 		else
1006 			push_out("%-9.9s ", dev_name);
1007 	}
1008 	col = 0;
1009 
1010 	knp = KSTAT_NAMED_PTR(&disk->is_errors);
1011 	for (i = 0; i < disk->is_errors.ks_ndata; i++) {
1012 		/* skip kstats that the driver did not kstat_named_init */
1013 		if (knp[i].name[0] == 0)
1014 			continue;
1015 
1016 		col += strlen(knp[i].name);
1017 
1018 		switch (knp[i].data_type) {
1019 			case KSTAT_DATA_CHAR:
1020 			case KSTAT_DATA_STRING:
1021 				if ((strcmp(knp[i].name, "Serial No") == 0) &&
1022 				    do_devid) {
1023 					if (disk->is_devid) {
1024 						push_out("Device Id: %s ",
1025 						    disk->is_devid);
1026 						col += strlen(disk->is_devid);
1027 					} else {
1028 						push_out("Device Id: ");
1029 					}
1030 
1031 					break;
1032 				}
1033 				if (knp[i].data_type == KSTAT_DATA_CHAR) {
1034 					push_out("%s: %-.16s ", knp[i].name,
1035 					    &knp[i].value.c[0]);
1036 					col += strnlen(&knp[i].value.c[0], 16);
1037 				} else {
1038 					push_out("%s: %s ", knp[i].name,
1039 					    KSTAT_NAMED_STR_PTR(&knp[i]));
1040 					col +=
1041 					    KSTAT_NAMED_STR_BUFLEN(&knp[i]) - 1;
1042 				}
1043 				break;
1044 			case KSTAT_DATA_ULONG:
1045 				push_out("%s: %u ", knp[i].name,
1046 				    knp[i].value.ui32);
1047 				col += 4;
1048 				break;
1049 			case KSTAT_DATA_ULONGLONG:
1050 				if (strcmp(knp[i].name, "Size") == 0) {
1051 					do_newline();
1052 					push_out("%s: %2.2fGB <%llu bytes>",
1053 					    knp[i].name,
1054 					    (float)knp[i].value.ui64 /
1055 					    DISK_GIGABYTE,
1056 					    knp[i].value.ui64);
1057 					do_newline();
1058 					col = 0;
1059 					break;
1060 				}
1061 				push_out("%s: %u ", knp[i].name,
1062 				    knp[i].value.ui32);
1063 				col += 4;
1064 				break;
1065 			}
1066 		if ((col >= 62) || (i == 2)) {
1067 			do_newline();
1068 			col = 0;
1069 		}
1070 	}
1071 	if (col > 0) {
1072 		do_newline();
1073 	}
1074 	do_newline();
1075 }
1076 
1077 void
1078 do_args(int argc, char **argv)
1079 {
1080 	int 		c;
1081 	int 		errflg = 0;
1082 	extern char 	*optarg;
1083 	extern int 	optind;
1084 
1085 	while ((c = getopt(argc, argv, "tdDxXYCciIpPnmMeEszrT:l:")) != EOF)
1086 		switch (c) {
1087 		case 't':
1088 			do_tty++;
1089 			break;
1090 		case 'd':
1091 			do_disk |= DISK_OLD;
1092 			break;
1093 		case 'D':
1094 			do_disk |= DISK_NEW;
1095 			break;
1096 		case 'x':
1097 			do_disk |= DISK_EXTENDED;
1098 			break;
1099 		case 'X':
1100 			if (do_disk & DISK_IOPATH_LTI)
1101 				errflg++;	/* -Y already used */
1102 			else
1103 				do_disk |= DISK_IOPATH_LI;
1104 			break;
1105 		case 'Y':
1106 			if (do_disk & DISK_IOPATH_LI)
1107 				errflg++;	/* -X already used */
1108 			else
1109 				do_disk |= DISK_IOPATH_LTI;
1110 			break;
1111 		case 'C':
1112 			do_controller++;
1113 			break;
1114 		case 'c':
1115 			do_cpu++;
1116 			break;
1117 		case 'I':
1118 			do_interval++;
1119 			break;
1120 		case 'p':
1121 			do_partitions++;
1122 			break;
1123 		case 'P':
1124 			do_partitions_only++;
1125 			break;
1126 		case 'n':
1127 			do_conversions++;
1128 			break;
1129 		case 'M':
1130 			do_megabytes++;
1131 			break;
1132 		case 'e':
1133 			do_disk |= DISK_ERRORS;
1134 			break;
1135 		case 'E':
1136 			do_disk |= DISK_EXTENDED_ERRORS;
1137 			break;
1138 		case 'i':
1139 			do_devid = 1;
1140 			break;
1141 		case 's':
1142 			suppress_state = 1;
1143 			break;
1144 		case 'z':
1145 			suppress_zero = 1;
1146 			break;
1147 		case 'm':
1148 			show_mountpts = 1;
1149 			break;
1150 		case 'T':
1151 			if (optarg) {
1152 				if (*optarg == 'u')
1153 					timestamp_fmt = UDATE;
1154 				else if (*optarg == 'd')
1155 					timestamp_fmt = DDATE;
1156 				else
1157 					errflg++;
1158 			} else {
1159 				errflg++;
1160 			}
1161 			break;
1162 		case 'r':
1163 			do_raw = 1;
1164 			break;
1165 		case 'l':
1166 			df.if_max_iodevs = safe_strtoi(optarg, "invalid limit");
1167 			if (df.if_max_iodevs < 1)
1168 				usage();
1169 			break;
1170 		case '?':
1171 			errflg++;
1172 	}
1173 
1174 	if ((do_disk & DISK_OLD) && (do_disk & DISK_NEW)) {
1175 		(void) fprintf(stderr, "-d and -D are incompatible.\n");
1176 		usage();
1177 	}
1178 
1179 	if (errflg) {
1180 		usage();
1181 	}
1182 
1183 	/* if no output classes explicity specified, use defaults */
1184 	if (do_tty == 0 && do_disk == 0 && do_cpu == 0)
1185 		do_tty = do_cpu = 1, do_disk = DISK_OLD;
1186 
1187 	/*
1188 	 * multi-path options (-X, -Y) without a specific vertical
1189 	 * output format (-x, -e, -E) imply extended -x format
1190 	 */
1191 	if ((do_disk & (DISK_IOPATH_LI | DISK_IOPATH_LTI)) &&
1192 	    !(do_disk & PRINT_VERTICAL))
1193 		do_disk |= DISK_EXTENDED;
1194 
1195 	/*
1196 	 * If conflicting options take the preferred
1197 	 * -D and -x result in -x
1198 	 * -d or -D and -e or -E gives only whatever -d or -D was specified
1199 	 */
1200 	if ((do_disk & DISK_EXTENDED) && (do_disk & DISK_NORMAL))
1201 		do_disk &= ~DISK_NORMAL;
1202 	if ((do_disk & DISK_NORMAL) && (do_disk & DISK_ERROR_MASK))
1203 		do_disk &= ~DISK_ERROR_MASK;
1204 
1205 	/* nfs, tape, always shown */
1206 	df.if_allowed_types = IODEV_NFS | IODEV_TAPE;
1207 
1208 	/*
1209 	 * If limit == 0 then no command line limit was set, else if any of
1210 	 * the flags that cause unlimited disks were not set,
1211 	 * use the default of 4
1212 	 */
1213 	if (df.if_max_iodevs == 0) {
1214 		df.if_max_iodevs = DEFAULT_LIMIT;
1215 		df.if_skip_floppy = 1;
1216 		if (do_disk & (DISK_EXTENDED | DISK_ERRORS |
1217 		    DISK_EXTENDED_ERRORS)) {
1218 			df.if_max_iodevs = UNLIMITED_IODEVS;
1219 			df.if_skip_floppy = 0;
1220 		}
1221 	}
1222 	if (do_disk) {
1223 		size_t count = 0;
1224 		size_t i = optind;
1225 
1226 		while (i < argc && !isdigit(argv[i][0])) {
1227 			count++;
1228 			i++;
1229 		}
1230 
1231 		/*
1232 		 * "Note:  disks  explicitly  requested
1233 		 * are not subject to this disk limit"
1234 		 */
1235 		if ((count > df.if_max_iodevs) ||
1236 		    (count && (df.if_max_iodevs == UNLIMITED_IODEVS)))
1237 			df.if_max_iodevs = count;
1238 
1239 		df.if_names = safe_alloc(count * sizeof (char *));
1240 		(void) memset(df.if_names, 0, count * sizeof (char *));
1241 
1242 		df.if_nr_names = 0;
1243 		while (optind < argc && !isdigit(argv[optind][0]))
1244 			df.if_names[df.if_nr_names++] = argv[optind++];
1245 	}
1246 	if (optind < argc) {
1247 		interval = safe_strtoi(argv[optind], "invalid interval");
1248 		if (interval < 1)
1249 			fail(0, "invalid interval");
1250 		optind++;
1251 
1252 		if (optind < argc) {
1253 			iter = safe_strtoi(argv[optind], "invalid count");
1254 			if (iter < 1)
1255 				fail(0, "invalid count");
1256 			optind++;
1257 		}
1258 	}
1259 	if (interval == 0)
1260 		iter = 1;
1261 	if (optind < argc)
1262 		usage();
1263 }
1264 
1265 /*
1266  * Driver for doing the extended header formatting. Will produce
1267  * the function stack needed to output an extended header based
1268  * on the options selected.
1269  */
1270 
1271 void
1272 do_format(void)
1273 {
1274 	char	header[SMALL_SCRATCH_BUFLEN] = {0};
1275 	char 	ch;
1276 	char 	iosz;
1277 	const char    *fstr;
1278 
1279 	disk_header[0] = 0;
1280 	ch = (do_interval ? 'i' : 's');
1281 	iosz = (do_megabytes ? 'M' : 'k');
1282 	if (do_disk & DISK_ERRORS) {
1283 		if (do_raw == 0) {
1284 			(void) sprintf(header, "s/w h/w trn tot ");
1285 		} else
1286 			(void) sprintf(header, "s/w,h/w,trn,tot");
1287 	}
1288 	switch (do_disk & DISK_IO_MASK) {
1289 		case DISK_OLD:
1290 			if (do_raw == 0)
1291 				fstr = "%cp%c tp%c serv  ";
1292 			else
1293 				fstr = "%cp%c,tp%c,serv";
1294 			(void) snprintf(disk_header, sizeof (disk_header),
1295 			    fstr, iosz, ch, ch);
1296 			break;
1297 		case DISK_NEW:
1298 			if (do_raw == 0)
1299 				fstr = "rp%c wp%c util  ";
1300 			else
1301 				fstr = "%rp%c,wp%c,util";
1302 			(void) snprintf(disk_header, sizeof (disk_header),
1303 			    fstr, ch, ch);
1304 			break;
1305 		case DISK_EXTENDED:
1306 			/* This is -x option */
1307 			if (!do_conversions) {
1308 				/* without -n option */
1309 				if (do_raw == 0) {
1310 					/* without -r option */
1311 					(void) snprintf(disk_header,
1312 					    sizeof (disk_header),
1313 					    "%-*.*s    r/%c    w/%c   "
1314 					    "%cr/%c   %cw/%c wait actv  "
1315 					    "svc_t  %%%%w  %%%%b %s",
1316 					    iodevs_nl, iodevs_nl, "device",
1317 					    ch, ch, iosz, ch, iosz, ch, header);
1318 				} else {
1319 					/* with -r option */
1320 					(void) snprintf(disk_header,
1321 					    sizeof (disk_header),
1322 					    "device,r/%c,w/%c,%cr/%c,%cw/%c,"
1323 					    "wait,actv,svc_t,%%%%w,"
1324 					    "%%%%b%s%s",
1325 					    ch, ch, iosz, ch, iosz, ch,
1326 					    *header == '\0' ? "" : ",",
1327 					    header);
1328 					/*
1329 					 * if no -e flag, header == '\0...'
1330 					 * Ternary operator above is to prevent
1331 					 * trailing comma in full disk_header
1332 					 */
1333 				}
1334 			} else {
1335 				/* with -n option */
1336 				if (do_raw == 0) {
1337 					fstr = "    r/%c    w/%c   %cr/%c   "
1338 					    "%cw/%c wait actv wsvc_t asvc_t  "
1339 					    "%%%%w  %%%%b %sdevice";
1340 				} else {
1341 					fstr = "r/%c,w/%c,%cr/%c,%cw/%c,"
1342 					    "wait,actv,wsvc_t,asvc_t,"
1343 					    "%%%%w,%%%%b,%sdevice";
1344 					/*
1345 					 * if -rnxe, "tot" (from -e) and
1346 					 * "device" are run together
1347 					 * due to lack of trailing comma
1348 					 * in 'header'. However, adding
1349 					 * trailing comma to header at
1350 					 * its definition leads to prob-
1351 					 * lems elsewhere so it's added
1352 					 * here in this edge case -rnxe
1353 					 */
1354 					if (*header != '\0')
1355 						(void) strcat(header, ",");
1356 				}
1357 				(void) snprintf(disk_header,
1358 				    sizeof (disk_header),
1359 				    fstr, ch, ch, iosz, ch, iosz,
1360 				    ch, header);
1361 			}
1362 			break;
1363 		default:
1364 			break;
1365 	}
1366 
1367 	/* do DISK_ERRORS header (already added above for DISK_EXTENDED) */
1368 	if ((do_disk & DISK_ERRORS) &&
1369 	    ((do_disk & DISK_IO_MASK) != DISK_EXTENDED)) {
1370 		if (!do_conversions) {
1371 			if (do_raw == 0)
1372 				(void) snprintf(disk_header,
1373 				    sizeof (disk_header), "%-*.*s  %s",
1374 				    iodevs_nl, iodevs_nl, "device", header);
1375 			else
1376 				(void) snprintf(disk_header,
1377 				    sizeof (disk_header), "device,%s", header);
1378 		} else {
1379 			if (do_raw == 0) {
1380 				(void) snprintf(disk_header,
1381 				    sizeof (disk_header),
1382 				    "  %sdevice", header);
1383 			} else {
1384 				(void) snprintf(disk_header,
1385 				    sizeof (disk_header),
1386 				    "%s,device", header);
1387 			}
1388 		}
1389 	} else {
1390 		/*
1391 		 * Need to subtract two characters for the % escape in
1392 		 * the string.
1393 		 */
1394 		dh_len = strlen(disk_header) - 2;
1395 	}
1396 
1397 	/*
1398 	 * -n *and* (-E *or* -e *or* -x)
1399 	 */
1400 	if (do_conversions && (do_disk & PRINT_VERTICAL)) {
1401 		if (do_tty)
1402 			setup(print_tty_hdr1);
1403 		if (do_cpu)
1404 			setup(print_cpu_hdr1);
1405 		if (do_tty || do_cpu)
1406 			setup(do_newline);
1407 		if (do_tty)
1408 			setup(print_tty_hdr2);
1409 		if (do_cpu)
1410 			setup(print_cpu_hdr2);
1411 		if (do_tty || do_cpu)
1412 			setup(do_newline);
1413 		if (do_tty)
1414 			setup(print_tty_data);
1415 		if (do_cpu)
1416 			setup(print_cpu_data);
1417 		if (do_tty || do_cpu)
1418 			setup(do_newline);
1419 		printxhdr();
1420 
1421 		setup(show_all_disks);
1422 	} else {
1423 		/*
1424 		 * These unholy gymnastics are necessary to place CPU/tty
1425 		 * data to the right of the disks/errors for the first
1426 		 * line in vertical mode.
1427 		 */
1428 		if (do_disk & PRINT_VERTICAL) {
1429 			printxhdr();
1430 
1431 			setup(show_first_disk);
1432 			if (do_tty)
1433 				setup(print_tty_data);
1434 			if (do_cpu)
1435 				setup(print_cpu_data);
1436 			setup(do_newline);
1437 
1438 			setup(show_other_disks);
1439 		} else {
1440 			setup(hdrout);
1441 			if (do_tty)
1442 				setup(print_tty_data);
1443 			setup(show_all_disks);
1444 			if (do_cpu)
1445 				setup(print_cpu_data);
1446 		}
1447 
1448 		setup(do_newline);
1449 	}
1450 	if (do_disk & DISK_EXTENDED_ERRORS)
1451 		setup(disk_errors);
1452 }
1453 
1454 /*
1455  * Add a new function to the list of functions
1456  * for this invocation. Once on the stack the
1457  * function is never removed nor does its place
1458  * change.
1459  */
1460 void
1461 setup(void (*nfunc)(void))
1462 {
1463 	format_t *tmp;
1464 
1465 	tmp = safe_alloc(sizeof (format_t));
1466 	tmp->nfunc = nfunc;
1467 	tmp->next = 0;
1468 	if (formatter_end)
1469 		formatter_end->next = tmp;
1470 	else
1471 		formatter_list = tmp;
1472 	formatter_end = tmp;
1473 
1474 }
1475 
1476 /*
1477  * The functions after this comment are devoted to printing
1478  * various parts of the header. They are selected based on the
1479  * options provided when the program was invoked. The functions
1480  * are either directly invoked in printhdr() or are indirectly
1481  * invoked by being placed on the list of functions used when
1482  * extended headers are used.
1483  */
1484 void
1485 print_tty_hdr1(void)
1486 {
1487 	char *fstr;
1488 	char *dstr;
1489 
1490 	if (do_raw == 0) {
1491 		fstr = "%10.10s";
1492 		dstr = "tty    ";
1493 	} else {
1494 		fstr = "%s";
1495 		dstr = "tty";
1496 	}
1497 	push_out(fstr, dstr);
1498 }
1499 
1500 void
1501 print_tty_hdr2(void)
1502 {
1503 	if (do_raw == 0)
1504 		push_out("%-10.10s", " tin tout");
1505 	else
1506 		push_out("tin,tout");
1507 }
1508 
1509 void
1510 print_cpu_hdr1(void)
1511 {
1512 	char *dstr;
1513 
1514 	if (do_raw == 0)
1515 		dstr = "     cpu";
1516 	else
1517 		dstr = "cpu";
1518 	push_out(dstr);
1519 }
1520 
1521 void
1522 print_cpu_hdr2(void)
1523 {
1524 	char *dstr;
1525 
1526 	if (do_raw == 0)
1527 		dstr = " us sy dt id";
1528 	else
1529 		dstr = "us,sy,dt,id";
1530 	push_out(dstr);
1531 }
1532 
1533 /*
1534  * Assumption is that tty data is always first - no need for raw mode leading
1535  * comma.
1536  */
1537 void
1538 print_tty_data(void)
1539 {
1540 	char *fstr;
1541 	uint64_t deltas;
1542 	double raw;
1543 	double outch;
1544 	kstat_t *oldks = NULL;
1545 
1546 	if (oldss)
1547 		oldks = &oldss->s_sys.ss_agg_sys;
1548 
1549 	if (do_raw == 0)
1550 		fstr = " %3.0f %4.0f ";
1551 	else
1552 		fstr = "%.0f,%.0f";
1553 	deltas = kstat_delta(oldks, &newss->s_sys.ss_agg_sys, "rawch");
1554 	raw = deltas;
1555 	raw /= getime;
1556 	deltas = kstat_delta(oldks, &newss->s_sys.ss_agg_sys, "outch");
1557 	outch = deltas;
1558 	outch /= getime;
1559 	push_out(fstr, raw, outch);
1560 }
1561 
1562 /*
1563  * Write out CPU data
1564  */
1565 void
1566 print_cpu_data(void)
1567 {
1568 	char *fstr;
1569 	uint64_t idle;
1570 	uint64_t user;
1571 	uint64_t kern;
1572 	uint64_t dtrace;
1573 	uint64_t nsec_elapsed;
1574 	kstat_t *oldks = NULL;
1575 
1576 	if (oldss)
1577 		oldks = &oldss->s_sys.ss_agg_sys;
1578 
1579 	if (do_raw == 0)
1580 		fstr = " %2.0f %2.0f %2.0f %2.0f";
1581 	else
1582 		fstr = "%.0f,%.0f,%.0f,%.0f";
1583 
1584 	idle = kstat_delta(oldks, &newss->s_sys.ss_agg_sys, "cpu_ticks_idle");
1585 	user = kstat_delta(oldks, &newss->s_sys.ss_agg_sys, "cpu_ticks_user");
1586 	kern = kstat_delta(oldks, &newss->s_sys.ss_agg_sys, "cpu_ticks_kernel");
1587 	dtrace = kstat_delta(oldks, &newss->s_sys.ss_agg_sys,
1588 	    "cpu_nsec_dtrace");
1589 	nsec_elapsed = newss->s_sys.ss_agg_sys.ks_snaptime -
1590 	    (oldks == NULL ? 0 : oldks->ks_snaptime);
1591 	push_out(fstr, user * percent, kern * percent,
1592 	    dtrace * 100.0 / nsec_elapsed / newss->s_nr_active_cpus,
1593 	    idle * percent);
1594 }
1595 
1596 /*
1597  * Emit the appropriate header.
1598  */
1599 void
1600 hdrout(void)
1601 {
1602 	if (do_raw == 0) {
1603 		if (--tohdr == 0)
1604 			printhdr(0);
1605 	} else if (hdr_out == 0) {
1606 		printhdr(0);
1607 		hdr_out = 1;
1608 	}
1609 }
1610 
1611 /*
1612  * Write out disk errors when -E is specified.
1613  */
1614 void
1615 disk_errors(void)
1616 {
1617 	(void) snapshot_walk(SNAP_IODEVS, oldss, newss, show_disk_errors, NULL);
1618 }
1619 
1620 void
1621 show_first_disk(void)
1622 {
1623 	int count = 0;
1624 
1625 	show_disk_mode = SHOW_FIRST_ONLY;
1626 
1627 	(void) snapshot_walk(SNAP_IODEVS, oldss, newss, show_disk, &count);
1628 }
1629 
1630 void
1631 show_other_disks(void)
1632 {
1633 	int count = 0;
1634 
1635 	show_disk_mode = SHOW_SECOND_ONWARDS;
1636 
1637 	(void) snapshot_walk(SNAP_IODEVS, oldss, newss, show_disk, &count);
1638 }
1639 
1640 void
1641 show_all_disks(void)
1642 {
1643 	int count = 0;
1644 
1645 	show_disk_mode = SHOW_ALL;
1646 
1647 	(void) snapshot_walk(SNAP_IODEVS, oldss, newss, show_disk, &count);
1648 }
1649 
1650 /*
1651  * Write a newline out and clear the lineout flag.
1652  */
1653 static void
1654 do_newline(void)
1655 {
1656 	if (lineout) {
1657 		(void) putchar('\n');
1658 		lineout = 0;
1659 	}
1660 }
1661 
1662 /*
1663  * Generalized printf function that determines what extra
1664  * to print out if we're in raw mode. At this time we
1665  * don't care about errors.
1666  */
1667 static void
1668 push_out(const char *message, ...)
1669 {
1670 	va_list args;
1671 
1672 	va_start(args, message);
1673 	if (do_raw && lineout == 1)
1674 		(void) putchar(',');
1675 	(void) vprintf(message, args);
1676 	va_end(args);
1677 	lineout = 1;
1678 }
1679 
1680 /*
1681  * Emit the header string when -e is specified.
1682  */
1683 static void
1684 print_err_hdr(void)
1685 {
1686 	char obuf[SMALL_SCRATCH_BUFLEN];
1687 
1688 	if (do_raw) {
1689 		push_out("errors");
1690 		return;
1691 	}
1692 
1693 	if (do_conversions == 0) {
1694 		if (!(do_disk & DISK_EXTENDED)) {
1695 			(void) snprintf(obuf, sizeof (obuf),
1696 			    "%11s", one_blank);
1697 			push_out(obuf);
1698 		}
1699 	} else if (do_disk == DISK_ERRORS)
1700 		push_out(two_blanks);
1701 	else
1702 		push_out(one_blank);
1703 	push_out("---- errors --- ");
1704 }
1705 
1706 /*
1707  * Emit the header string when -e is specified.
1708  */
1709 static void
1710 print_disk_header(void)
1711 {
1712 	push_out(disk_header);
1713 }
1714 
1715 /*
1716  * No, UINTMAX_MAX isn't the right thing here since
1717  * it is #defined to be either INT32_MAX or INT64_MAX
1718  * depending on the whether _LP64 is defined.
1719  *
1720  * We want to handle the odd future case of having
1721  * ulonglong_t be more than 64 bits but we have
1722  * no nice #define MAX value we can drop in place
1723  * without having to change this code in the future.
1724  */
1725 
1726 u_longlong_t
1727 ull_delta(u_longlong_t old, u_longlong_t new)
1728 {
1729 	if (new >= old)
1730 		return (new - old);
1731 	else
1732 		return ((UINT64_MAX - old) + new + 1);
1733 }
1734 
1735 /*
1736  * Take the difference of an unsigned 32
1737  * bit int attempting to cater for
1738  * overflow.
1739  */
1740 uint_t
1741 u32_delta(uint_t old, uint_t new)
1742 {
1743 	if (new >= old)
1744 		return (new - old);
1745 	else
1746 		return ((UINT32_MAX - old) + new + 1);
1747 }
1748 
1749 /*
1750  * This is exactly what is needed for standard iostat output,
1751  * but make sure to use it only for that
1752  */
1753 #define	EPSILON	(0.1)
1754 static int
1755 fzero(double value)
1756 {
1757 	return (value >= 0.0 && value < EPSILON);
1758 }
1759 
1760 static int
1761 safe_strtoi(char const *val, char *errmsg)
1762 {
1763 	char *end;
1764 	long tmp;
1765 
1766 	errno = 0;
1767 	tmp = strtol(val, &end, 10);
1768 	if (*end != '\0' || errno)
1769 		fail(0, "%s %s", errmsg, val);
1770 	return ((int)tmp);
1771 }
1772